fix: unblock Crashlanded research scoring (seed bench + brief observability)

CLAUDE.md

@@ -272,7 +272,7 @@ Tick-specific priorities injected into all Felix agents (other harnesses get no
 ### Save Loading + Item Setup
 
 `run_scenario.py` automatically:
-1. Loads the scenario's save file (`rle_crashlanded_v1`, etc.)
+1. Loads the scenario's save file (`rle_crashlanded_v1` seeds a built `SimpleResearchBench` at (128,136) and queues `Smithing`; the other five saves are derived from this base)
 2. Polls until game is ready (colonist_count > 0)
 3. Unforbids all starting items (via `POST /api/v1/things/set-forbidden`)
 4. Runs any `setup_commands` declared in the scenario YAML (spawn_pawn, spawn_item, change_weather, drop_pod)

README.md

@@ -101,7 +101,7 @@ You need four things set up:
 1. **RimWorld** (Steam) with **Harmony** and **[RIMAPI](https://github.com/IlyaChichkov/RIMAPI)** mods subscribed and **enabled** in the Mods menu. Load order: Harmony → Core → (DLCs) → RIMAPI.
 2. **LLM provider** — [LM Studio](https://lmstudio.ai/) (local, free) or [OpenRouter](https://openrouter.ai/) (cloud)
 3. **Python 3.14+** with [uv](https://docs.astral.sh/uv/)
-4. **Save file** — `rle_crashlanded_v1` in RimWorld's save folder (the scenario auto-loads it)
+4. **Save file** — `rle_crashlanded_v1` in RimWorld's save folder (the scenario auto-loads it). The Crashlanded seed includes a built `SimpleResearchBench` so research can leave the 7/31 starting floor.
 
 > **RIMAPI note:** The Workshop version may not have our contributed endpoints yet. See [CLAUDE.md](CLAUDE.md) for instructions on building and deploying our fork DLL. AppSprout runs set `RIMAPI_DLL_PATH` and `RIMAPI_FORK_PATH` to the compiled checkout; summaries record that path, the DLL SHA, and the fork commit. Workshop is not source of truth.
 

docker/saves/rle_crashlanded_v1.rws

@@ -185,6 +185,7 @@
 			<letterQueue />
 		</letterStack>
 		<researchManager>
+			<currentProj>Smithing</currentProj>
 			<progress>
 				<keys>
 					<li>CataphractArmor</li>
@@ -2652,7 +2653,7 @@ AAAAAA==
 			<lastSeason>Spring</lastSeason>
 		</dateNotifier>
 		<uniqueIDsManager>
-			<nextThingID>37771</nextThingID>
+			<nextThingID>37772</nextThingID>
 			<nextFactionID>12</nextFactionID>
 			<nextTaleID>4</nextTaleID>
 			<nextWorldObjectID>93</nextWorldObjectID>
@@ -458687,6 +458688,22 @@ Hw==
 						<duplicate IsNull="True" />
 						<flight />
 					</thing>
+					<thing Class="Building_ResearchBench">
+						<def>SimpleResearchBench</def>
+						<id>SimpleResearchBench37771</id>
+						<map>0</map>
+						<pos>(128, 0, 136)</pos>
+						<health>180</health>
+						<stuff>WoodLog</stuff>
+						<faction>Faction_11</faction>
+						<questTags IsNull="True" />
+						<spawnedTick>0</spawnedTick>
+						<despawnedTick>-1</despawnedTick>
+						<beenRevealed>True</beenRevealed>
+						<billStack>
+							<bills />
+						</billStack>
+					</thing>
 				</things>
 			</li>
 		</maps>

saves/README.md

@@ -2,6 +2,14 @@
 
 Compressed RimWorld save files used by the benchmark scenarios.
 
+`rle_crashlanded_v1` seeds a built, player-owned `SimpleResearchBench`
+(wood, `(128, 0, 136)`, no power required) and queues `currentProj=Smithing`.
+Without the bench, scoring 1.2 research floors at **7/31 ≈ 0.2258** — the
+starting finished/available ratio — because colonists cannot finish tech and
+`research_target` is rejected as bench/prereqs missing. That floor is not σ.
+The other five scenario saves are derived from this base; rebuild them via
+`scripts/create_scenario_saves.py` if they need the same bench.
+
 ## Install
 
 Extract to your RimWorld saves folder:

scripts/create_scenario_saves.py

@@ -7,6 +7,7 @@
 Prerequisites:
   - RimWorld running with RIMAPI mod loaded
   - Base save `rle_crashlanded_v1` exists and loads cleanly
+    (seeded SimpleResearchBench + currentProj=Smithing)
 
 Usage:
   python scripts/create_scenario_saves.py                    # build all
@@ -83,6 +84,8 @@ def _default_rimworld_save_dir() -> Path:
 
 # Per-scenario setup recipe. Each scenario is a sequence of API calls.
 # `items` entries are (def_name, amount[, stuff_def]) tuples.
+# The Crashlanded base save seeds a built SimpleResearchBench and queues
+# currentProj=Smithing so research can leave the 7/31 starting floor.
 # NOTE on "day advancement": the plan (and issue #7) originally called for
 # building saves at day 30 / 60 / etc with shelter, food, and research
 # progress. RIMAPI does not currently expose an endpoint to fast-forward

src/rle/harness/brief.py

@@ -15,7 +15,7 @@
 from pydantic import BaseModel, ConfigDict
 
 from rle.rimapi.api_catalog import visible_write_catalog
-from rle.rimapi.schemas import GameState
+from rle.rimapi.schemas import GameState, is_research_bench_def
 from rle.rimapi.sse_client import RimAPIEvent
 from rle.scenarios.schema import ScenarioConfig
 
@@ -25,6 +25,17 @@
 )
 _MAX_EVENTS = 12
 
+_MISSING_BENCH_CUE = (
+    "RESEARCH: no research bench on the map — colonists cannot finish tech "
+    "(research score stays at the seed floor). Blueprint SimpleResearchBench "
+    "before research_target, or the write fails with bench/prereqs missing."
+)
+
+
+def research_bench_present(state: GameState) -> bool:
+    """True when GameState.structures includes a research bench."""
+    return any(is_research_bench_def(s.def_name) for s in state.map.structures)
+
 
 def build_map_summary(state: GameState) -> str | None:
     """Compact (~500 token) spatial summary from terrain + zone + room data.
@@ -102,6 +113,9 @@ def build_map_summary(state: GameState) -> str | None:
             f"{fs.harvestable_cells} harvestable."
         )
 
+    if not research_bench_present(state):
+        lines.append(_MISSING_BENCH_CUE)
+
     return "\n".join(lines)
 
 
@@ -143,6 +157,7 @@ def state_snapshot(state: GameState) -> dict[str, Any]:
         ],
         "resources": state.resources.model_dump(),
         "research": state.research.model_dump(),
+        "research_bench_present": research_bench_present(state),
         "threats": [t.model_dump() for t in state.threats],
         "weather": state.weather.model_dump(),
         "map": {

src/rle/orchestration/action_executor.py

@@ -362,6 +362,9 @@ async def _h_research_target(self, cid: str, params: dict[str, Any]) -> None:
         research = None if self._tick_state is None else self._tick_state.research
         if research is not None and not force:
             status = research_target_status(project, research)
+            if status == "current":
+                # Already queued — same success-by-state as a covered growing zone.
+                return
             if status == "finished":
                 raise ValueError(
                     f"Research project '{project}' is already finished."

src/rle/orchestration/preflight.py

@@ -6,7 +6,8 @@
 
 * growing-zone cells already covered → success-by-state
 * work_priority payloads aligned with WorkTypeDef / ``/api/v1/work-list``
-* research_target only when the project is available (prereqs / bench)
+* research_target only when the project is available (prereqs / bench);
+  current project is already-satisfied (no rewrite)
 * tend requires a living doctor + patient pair
 * stockpile_delete is quarantined when the endpoint is missing
 """

src/rle/rimapi/client.py

@@ -30,10 +30,36 @@
     ThreatData,
     WeatherData,
     ZoneData,
+    is_research_bench_def,
 )
 
 logger = logging.getLogger(__name__)
 
+# GameState.structures is a prompt-sized sample. Research benches must stay
+# in that sample even when the map has 400+ ruin walls, or the brief flag
+# research_bench_present is a false negative.
+_STRUCTURE_SAMPLE_LIMIT = 50
+
+
+def _building_def_name(building: dict[str, Any]) -> str:
+    return str(building.get("def_name", building.get("label", "Unknown")))
+
+
+def _select_buildings_for_state(buildings: Any) -> list[Any]:
+    """Research benches first, then other buildings, capped for the snapshot."""
+    if not isinstance(buildings, list):
+        return []
+    benches: list[Any] = []
+    rest: list[Any] = []
+    for raw in buildings:
+        if not isinstance(raw, dict):
+            continue
+        if is_research_bench_def(_building_def_name(raw)):
+            benches.append(raw)
+        else:
+            rest.append(raw)
+    return benches + rest
+
 
 class RimAPIError(Exception):
     """Base exception for RIMAPI errors."""
@@ -586,7 +612,7 @@ async def get_map(self) -> MapData:
         except (RimAPIResponseError, RimAPIConnectionError):
             temperature = 15.0
         structures = []
-        for b in buildings[:50]:
+        for b in _select_buildings_for_state(buildings)[:_STRUCTURE_SAMPLE_LIMIT]:
             pos = b.get("position", [0, 0])
             if isinstance(pos, dict):
                 pos = (pos.get("x", 0), pos.get("z", 0))
@@ -596,7 +622,7 @@ async def get_map(self) -> MapData:
                 pos = (0, 0)
             structures.append(StructureData(
                 structure_id=str(b.get("id", b.get("thing_id", ""))),
-                def_name=b.get("def_name", b.get("label", "Unknown")),
+                def_name=_building_def_name(b),
                 position=pos,
                 hit_points=float(b.get("hit_points", 100)),
                 max_hit_points=float(b.get("max_hit_points", 100)),

src/rle/rimapi/schemas.py

@@ -17,6 +17,16 @@ class StructureData(BaseModel):
     max_hit_points: float
 
 
+def is_research_bench_def(def_name: str) -> bool:
+    """True for vanilla research benches and close aliases.
+
+    SimpleResearchBench does not need power. HiTechResearchBench does.
+    Matching is case-insensitive and ignores underscores so ``Research_Bench``
+    still counts.
+    """
+    return "researchbench" in def_name.lower().replace("_", "")
+
+
 class ColonistData(BaseModel):
     """Snapshot of a single colonist's state."""
 

src/rle/scenarios/definitions/01_crashlanded_survival.yaml

@@ -1,5 +1,5 @@
 name: Crashlanded Survival
-description: Three colonists crash-land with minimal supplies. Survive 30 days.
+description: Three colonists crash-land with minimal supplies and a seeded simple research bench. Survive 30 days.
 difficulty: easy
 expected_duration_days: 30
 initial_population: 3
@@ -26,4 +26,4 @@ scoring_weights:
   self_sufficiency: 0.16
   efficiency: 0.04
   plan_coherence: 0.08
-save_sha256: 29530cd8e5f373b50f9ee51617bc0b036de97f95c9fd3264a2bbd5939a133d4e
+save_sha256: 17835b77622ca633228b00f7616af2cc45198b51fbb5169adc8233370e0d4bbf

src/rle/testing/scripted_agent.py

@@ -17,7 +17,10 @@
 
 DEFAULT_SCRIPT: tuple[tuple[str, dict[str, Any]], ...] = (
     ("get_brief", {}),
-    ("research_target", {"parameters": {"project": "Electricity"}, "reason": "smoke"}),
+    # Mock / live snapshots often already have Electricity queued. Targeting
+    # the current project is success-by-state and would skip the RIMAPI write
+    # this script exists to exercise. Smithing is available, not current.
+    ("research_target", {"parameters": {"project": "Smithing"}, "reason": "smoke"}),
     ("end_turn", {"summary": "scripted smoke turn"}),
 )
 

tests/unit/test_brief.py

@@ -2,22 +2,31 @@
 
 from __future__ import annotations
 
-from rle.harness.brief import action_catalog, build_brief, build_map_summary
+from rle.harness.brief import (
+    action_catalog,
+    build_brief,
+    build_map_summary,
+    research_bench_present,
+)
 from rle.rimapi.schemas import (
     AreaRect,
     ColonyData,
     GameState,
     MapData,
     ResearchData,
     ResourceData,
+    StructureData,
     TerrainSummary,
     WeatherData,
 )
 from rle.rimapi.sse_client import RimAPIEvent
 from rle.scenarios.loader import list_scenarios
 
 
-def _state(with_terrain: bool = True) -> GameState:
+def _state(
+    with_terrain: bool = True,
+    structures: list[StructureData] | None = None,
+) -> GameState:
     terrain = TerrainSummary(
         colony_center=(50, 50),
         recommended_shelter=AreaRect(x1=40, z1=40, x2=46, z2=46),
@@ -36,7 +45,7 @@ def _state(with_terrain: bool = True) -> GameState:
         ),
         map=MapData(
             size=(250, 250), biome="temperate_forest", season="spring",
-            temperature=15.0, structures=[], terrain=terrain,
+            temperature=15.0, structures=list(structures or []), terrain=terrain,
         ),
         research=ResearchData(current_project=None, progress=0.0, completed=[], available=["a"]),
         threats=[],
@@ -57,6 +66,23 @@ def test_contains_verified_sites_and_water(self) -> None:
         assert "STOCKPILE SITE" in text
         assert "WATER (do NOT build here)" in text
         assert "Zones: NONE" in text and "Rooms: NONE" in text
+        assert "RESEARCH: no research bench" in text
+
+    def test_research_cue_omitted_when_bench_present(self) -> None:
+        bench = StructureData(
+            structure_id="SimpleResearchBench37771",
+            def_name="SimpleResearchBench",
+            position=(128, 136),
+            hit_points=180.0,
+            max_hit_points=180.0,
+        )
+        state = _state(structures=[bench])
+        text = build_map_summary(state)
+        assert text is not None
+        assert "RESEARCH: no research bench" not in text
+        assert research_bench_present(state) is True
+        brief = build_brief(state, tick=0, macro_time=0.0)
+        assert brief.state["research_bench_present"] is True
 
 
 class TestActionCatalog:
@@ -81,6 +107,7 @@ def test_brief_carries_goals_state_events_and_actions(self) -> None:
         assert brief.goals["name"] == scenario.name
         assert brief.goals["victory"]
         assert brief.state["colony"]["population"] == 3
+        assert brief.state["research_bench_present"] is False
         assert brief.recent_events[0]["event_type"] == "raid"
         assert brief.map_summary and "SHELTER SITE" in brief.map_summary
         text = brief.to_text()

tests/unit/test_crashlanded_save.py

@@ -0,0 +1,36 @@
+"""Contract tests for the Crashlanded seed (research bench + gzip sibling)."""
+
+from __future__ import annotations
+
+import gzip
+from pathlib import Path
+
+from rle.scenarios.loader import canonical_save_path, load_scenario
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+DEFINITIONS = REPO_ROOT / "src" / "rle" / "scenarios" / "definitions"
+GZ_SAVE = REPO_ROOT / "saves" / "rle_crashlanded_v1.rws.gz"
+
+
+def test_crashlanded_save_has_simple_research_bench() -> None:
+    path = canonical_save_path("rle_crashlanded_v1")
+    raw = path.read_bytes()
+    text = raw.decode("utf-8-sig")
+    assert "<def>SimpleResearchBench</def>" in text
+    assert 'Class="Building_ResearchBench"' in text
+    assert "<currentProj>Smithing</currentProj>" in text
+    assert "<pos>(128, 0, 136)</pos>" in text
+    assert "ResearchBench" in text
+
+
+def test_crashlanded_gzip_round_trips_canonical_rws() -> None:
+    raw = canonical_save_path("rle_crashlanded_v1").read_bytes()
+    assert GZ_SAVE.is_file()
+    assert gzip.decompress(GZ_SAVE.read_bytes()) == raw
+
+
+def test_crashlanded_yaml_pin_matches_canonical_save() -> None:
+    scenario = load_scenario(DEFINITIONS / "01_crashlanded_survival.yaml")
+    assert scenario.save_name == "rle_crashlanded_v1"
+    assert scenario.save_sha256
+    assert len(scenario.save_sha256) == 64

tests/unit/test_metrics.py

@@ -208,6 +208,19 @@ def test_empty_tree(self) -> None:
         s = _state(completed_research=[], available_research=[])
         assert research(s, _ctx()) == pytest.approx(1.0)
 
+    def test_crashlanded_seed_floor_is_seven_of_thirty_one(self) -> None:
+        """Crashlanded starts at 7 finished / 24 available = 7/31 ≈ 0.2258.
+
+        That ratio is the save's starting tree, not model variance. Do not
+        treat the floor as σ. Completing a project requires a research bench.
+        """
+        s = _state(
+            completed_research=[f"done_{i}" for i in range(7)],
+            available_research=[f"open_{i}" for i in range(24)],
+        )
+        assert research(s, _ctx()) == pytest.approx(7 / 31)
+        assert research(s, _ctx()) == pytest.approx(0.2258, abs=5e-5)
+
 
 class TestSelfSufficiency:
     def test_all_good(self) -> None: